YOLOv11-SFF: An Improved Algorithm for Detecting Foreign Objects on Mine Conveyor Belts

Abstract

To solve the problem of detecting foreign objects on mine conveyor belts with different backgrounds and variations in target scales, difficulty recognising elongated targets, and missed targets or false negatives for small foreign objects, this paper proposes YOLOv11-SFF, a conveyor belt foreign object detection algorithm based on an improved version of YOLOv11n. YOLOv11-SFF first introduces a C3k2_S module based on direction-aware strip convolutions to enhance the model’s ability to learn the long-term dependence structure of thin targets. Strip convolution performs horizontal and vertical convolution operations. Secondly, there is an FSPPF module that combines the FReLU activation function to augment deep feature representation and aggregate multi-scale contextual information for improved recognition, enabling the model to detect small objects. Finally, the FAFM Feature Alignment and Fusion Module is added in the neck network. Deformable convolutions and content-aware attention are used to achieve cross-scale feature space alignment and adaptive fusion, thus improving the detection accuracy of small, long and low-contrast objects on the conveyor belt. Experimental results demonstrate that, compared to the baseline model, the precision, mAP@50 and mAP@50-95 of YOLOv11-SFF increase by 2.4, 2.6 and 1.9 percentage points respectively, validating the effectiveness of YOLOv11-SFF in the task of detecting foreign objects on mine conveyor belts.